Zero-knowledge machine learning

We ZK-SNARK the first ImageNet model for fully privacy preserving model weights or image inputs.

To do so, we leverage advances in ZK technology and build a conversion method for taking trained ML models and transforming them into ZK-SNARKs. Our method runs on commodity hardware.

We use halo2, tensorflow, and rust to build our project. We built a DSL for converting tensorflow models into halo2.

The DSL is built in rust. Our DSL currently supports vision models. We're planning to expand to transformers and other kinds of models in the future as well.

For parallelization we use rayon.